Liam Cawley
Investigating whether transformer in-context learning implements kernel ridge regression on learned hidden representations. We construct explicit linear-attention-to-conjugate-gradient mappings and study the softmax extension.
Liam Cawley
A toolkit for computing ε-Rashomon sets, membership certificates, and set-level interpretability metrics for GLMs. Addresses the question: when many models fit the data equally well, which explanations are stable?
Hugh van Deventer, Itamar Pres, Liam Cawley
A benchmark for evaluating steering methods that operate through a language model's unembedding matrix. Provides standardized comparisons across extraction methods, models, and tasks.
Liam Cawley
An undergraduate project (Jun 2024) framing LoRA rank allocation as a budgeted resource-allocation problem, with offline curvature-aware and online greedy algorithms. Added in retrospect with notes on where the original claims outrun the evidence.
Liam Cawley
A minimal distributed training sandbox for nanoGPT. Experiments with MIG partitioning, NCCL collectives, Kubernetes orchestration, and mixed-precision training on small-scale hardware.
Liam Cawley, Alexandra Lavacek, Sophia Tesic
Adapting denoising diffusion concepts to single-image super resolution. We progressively build from a naive upsampler to a residual architecture with channel attention and perceptual loss, achieving 34.0 dB PSNR on DIV2K at 2x upscaling.
Liam Cawley
An undergraduate benchmark for conversation-history poisoning attacks on language models, covering false conversation injection, gaslighting, and iterative context poisoning. Added in retrospect with notes on what the setup does and does not establish.
Liam Cawley, Gabe Ronan
A closed-loop calibration application for an active electronically scanned phased-array radar, optimizing 6-bit phase-shifter and attenuator states on Anokiwave AWS-0103 beamformers with metaheuristic search and a CNN surrogate, reaching a 6.5x reduction in beam-synthesis error over a Sample Matrix Inversion baseline.
Liam Cawley
An empirical comparison of stochastic regularization methods --- Shake-Shake, Mixup, and Cutout --- in residual networks on CIFAR-10, with analysis of when and why each technique helps.
Liam Cawley
Design of a fully autonomous fixed-wing UAV on the MFD Crosswind platform for agricultural seed spreading and mapping, benchmarked against the AUVSI SUAS 2021 mission of waypoint navigation, object detection, and a parachute UGV airdrop.